In short: Most of what DDMRP delivers comes from decoupling, which shortens the horizon any buffer has to cover, and rather less comes from the buffer formula itself, whose lead time and variability factors are categorical bands a planner picks from a table. Decode a default buffer on a single item and it implies a protection level around 99.99%, roughly double the safety stock a 95% cycle service target would have asked for on the same inputs. That makes any comparison against an MRP running zero safety stock and a full cumulative lead time a measurement of parameterisation. Miclo and colleagues (2019) and Kortabarria and colleagues (2018) both compare by simulation on single settings, and neither tunes the MRP arm.
A supplier confirms one purchased component fourteen days later than promised. The overnight regeneration takes that date and pushes it through nine parents, and by Tuesday morning there are roughly three hundred rescheduled lines waiting for someone to look at. Most of them will be reversed within a fortnight. The planner works the list top down until lunchtime, gets through maybe forty, and the other two hundred and sixty are decided by not being looked at.
The reflex is to blame MRP for that morning, and the mechanics of why regeneration behaves this way belong with MM1. What matters here is that demand driven MRP is usually sold as the fix for it, and the sales conversation moves quickly to coloured buffers. The buffers are the least interesting part of the method.
Decoupling is the mechanism
Take a finished pack with three levels underneath it. Resin bought from an overseas supplier on a 35 day lead time. An in-house compounding step at 6 days. Fill and pack at 3 days. Cumulative lead time through that structure is 44 days, and any promise you make on the finished item is exposed to every source of variability across all 44 of them.
Now hold stock deliberately at the compound, meaning a position you chose to buffer and manage rather than one that accumulated because a forecast was wrong. The decoupled lead time for the finished item becomes 3 days, because the compound is there when fill and pack asks for it. Buffer the resin as well and the compound's own decoupled lead time drops from 41 days to 6.
That is the lead time compression the method claims, and it is real. It is also entirely a stock positioning decision. You have converted a 44 day exposure on the item you quote into a 3 day exposure, and you have paid for it with inventory at two positions further back. The compression is a transfer, and the reason it is usually a good transfer is that variability is cheaper to absorb at a compound than at a finished pack with forty variants.
Nothing in that paragraph requires DDMRP software. A planner running conventional MRP who stocks the compound and the resin, sets a planning time fence on the finished item and plans the lower levels to a buffer gets the same structural effect. What DDMRP adds is a stated procedure for choosing the positions, a buffer that resizes as usage moves, and a signal a planner can act on without reading a rescheduling list.
Choosing the positions is where the judgement sits. The candidates are places where lead time is long and variable below and short above, where a component is shared across many parents, where a customer tolerance time is shorter than the cumulative lead time, and where a process step creates variety from commonality. Those criteria conflict on most real bills of material, and there is no formula that resolves them. Where the buffer should sit across a distribution network is a different question with its own literature, and that one belongs with I1.
What the buffer zones actually compute
The zone arithmetic is simple enough to work by hand, which is worth doing once before you accept it.
Take an item with average daily usage of 100 units and a decoupled lead time of 20 days. Ptak and Smith's buffer profiles assign it a lead time factor, from a band running roughly 0.20 to 0.40 for long lead times, and a variability factor from a band running roughly 0.41 to 0.60 for medium variability. Say 0.40 and 0.50.
The red zone base is usage times lead time times the lead time factor, so 100 times 20 times 0.40, which is 800. Red safety is the base times the variability factor, 400. Top of red is 1,200. The yellow zone is usage times lead time, 2,000, so top of yellow is 3,200. Green is the largest of the minimum order quantity, the base calculation of 800, and usage times any imposed order cycle. Call it 800. Top of green is 4,000.
Net flow position is on hand plus on order minus qualified demand. When it falls into yellow you order up to top of green. Expected average on hand is top of red plus half the green zone, so 1,200 plus 400, or 1,600 units. Sixteen days of cover.
Two numbers set all of that: the lead time factor and the variability factor. Both are chosen from a small set of bands by a person, and in most implementations they are chosen once, by profile, for a whole class of items. DDMRP is often positioned as removing a parameter nobody tunes, which is fair as a description of what safety stock looks like in most ERP systems. It introduces two parameters nobody tunes either, and the tuning is categorical rather than measured.
Comparing against MRP carrying the same stock
Give the same item a demand standard deviation of 30 units a day and a supplier whose lead time has a standard deviation of 3 days. The variance of demand over a random lead time has two terms, one for demand variability and one for lead time variability. Twenty times 900 is 18,000. Mean demand squared times lead time variance is 10,000 times 9, or 90,000. Total 108,000, standard deviation 329 units. The sizing of that expression, including the assumptions it makes, is I2's subject.
At a 95% cycle service level the safety stock is 1.65 times 329, or 542 units. With the same order quantity of 800, average on hand is 542 plus 400, so 942 units. Nine and a half days.
The DDMRP buffer on identical inputs holds 1,600. Read top of red as the protection level, and 1,200 units against a standard deviation of 329 is a z of 3.65, which corresponds to a cycle service level of about 99.99%. Nobody chose that. It fell out of two categorical factors picked from a table.
Neither number is right by construction. The point of the arithmetic is that the two methods are separated by parameter choices, and the parameters are visible on both sides. A comparison that runs default DDMRP buffers against an MRP configured with zero safety stock and the full 44 day cumulative lead time is comparing 1,600 units of protection against roughly 400, and the result of that comparison was decided before the simulation ran.
A comparison worth acting on holds four things equal. Both arms get the same decoupling structure, meaning the same positions are stocked. Both get the same order quantities. Both get the same target service level, with the DDMRP factors solved backwards to hit it rather than taken from a profile. And both run over the same history with a rolling origin, so you see how each behaves through the periods where supply actually failed. Then compare average on hand and fill rate, and see what is left.
One parameter deserves specific attention in that test. Whybark and Williams showed in 1976 that under lead time uncertainty a safety lead time outperforms a safety stock of equivalent cost, while under quantity uncertainty the reverse holds. DDMRP folds lead time uncertainty into the lead time factor, which multiplies a quantity. An MRP arm given a safety lead time on the long pole component is being parameterised against the right failure mode, and it closes more of the gap than most bake-offs allow it to.
What the published evaluations establish
Miclo, Lauras, Fontanili, Lamothe and Melnyk published a simulation comparison in the International Journal of Production Research in 2019, setting DDMRP against MRP II and a lean flow approach, and reported DDMRP ahead on service and inventory in the configurations they tested. Kortabarria, Apaolaza, Lizarralde and Amorrortu ran a simulated changeover inside one industrial firm for the Journal of Industrial Engineering and Management in 2018 and reported lower inventory at held service.
Both are careful pieces of work and both share a limitation the method's marketing tends to drop. Each tests a small number of settings, and neither randomises the parameterisation of the comparison arm. They establish that DDMRP works. They do not establish what a well tuned MRP with the same decoupling structure would have done, because that arm was not run.
The method's own definition sits with Ptak and Smith, who rewrote Orlicky's 1975 text in its third edition in 2011 and published the DDMRP specification separately. Treating the specification as the authority on what the method is and the peer reviewed simulations as the evidence on how it performs keeps the two straight.
Where this stops
DDMRP is weakest exactly where conventional MRP is strongest. High volume repetitive manufacture, stable schedules, reliable supply and genuinely dependent demand are conditions where determinism beats buffering, and buffering every level in that environment adds stock to protect against variability that is not there.
Shared components have no clean answer. A component feeding forty parents with different customer tolerance times is a decoupling candidate by every criterion, and once you buffer it you have arrived at MRP with a safety stock, reached by a different route and described in different vocabulary.
The average daily usage window is a moving average, so it lags trend and seasonality in both directions. On a seasonal item the buffer grows after the season has started and shrinks after it has ended, and the correction is the demand adjustment factor, which is a planner's judgement applied to a number that was meant to remove judgement.
There is also a straightforward point about who is speaking. The specification, the certification and most of the favourable case material come from one organisation. That does not make the method wrong, and the peer reviewed simulations above are independent of it. It does mean the reported percentage improvements should be read as what a specific implementation achieved against a specific baseline, and not as what the method will do for you.
Take your ten highest volume finished items, write down the cumulative lead time through the bill of materials for each, then write down the longest path that has no stocked position anywhere on it. The gap between those two columns is the entire prize decoupling has to offer, and you can calculate it this week with a spreadsheet and your existing item master.